Dynamic Higher-Order Information Bottleneck with Adaptive Hypergraph Learning for Brain Disease Diagnosis

Traditional functional magnetic resonance imaging (fMRI) analysis often relies on static pairwise functional connectivity, overlooking the dynamic and higher-order interactions crucial for understanding complex brain disorders. To address this, we propose DynHOIB, a Dynamic Higher-Order Information Bottleneck framework with Adaptive Hypergraph Learning. DynHOIB dynamically captures multi-view information from fMRI time series, integrating both pairwise and higher-order hypergraph representations. It employs a learnable attention module for adaptive higher-order interaction modeling, an adaptive hypergraph learning component, and a Dynamic Hypergraph Neural Network to process evolving structures. A multi-level information bottleneck mechanism hierarchically distills the most discriminative features across temporal and view dimensions. Experiments on multiple fMRI datasets demonstrate that DynHOIB achieves superior classification performance and captures more clinically relevant and biologically interpretable higher-order brain interactions.